/
sft.py
149 lines (126 loc) · 5.51 KB
/
sft.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
from argparse import ArgumentParser
from collections import Counter
from pathlib import Path
from typing import Any
import torch
from collections import Counter
from datasets import Dataset, DatasetDict, load_dataset, concatenate_datasets
from peft import LoraConfig # type: ignore
from transformers import (
AutoModelForCausalLM,
AutoTokenizer,
DataCollatorForLanguageModeling,
TrainingArguments,
)
from trl import SFTTrainer
from train_utils import assert_type
class LastTokenOnlyDataCollator(DataCollatorForLanguageModeling):
def torch_call(
self, examples: list[dict[str, Any]]
) -> dict[str, Any]:
# pass only input_ids and attention_mask for super().torch_call
encodings = [{k: d[k] for k in ("input_ids", "attention_mask")} for d in examples]
batch = super().torch_call(encodings)
# Compute the sequence length of each sample in the batch
seq_lens = torch.sum(batch["input_ids"] != tokenizer.pad_token_id, dim=1)
# Create a new tensor for the labels, fill it with -100, then copy over
# only the last token for each sequence
old_labels = batch["labels"]
batch["labels"] = torch.full_like(old_labels, -100).scatter_(
1, seq_lens[:, None] - 1, old_labels.gather(1, seq_lens[:, None] - 1)
)
return batch
def balance(ds: Dataset) -> Dataset:
"""Balance a dataset by undersampling the majority class."""
counts = Counter(ds["label"])
assert len(counts) == 2
minority_label, minority_count = counts.most_common()[1]
majority_label, _ = counts.most_common()[0]
minority_ds = ds.filter(lambda x: x["label"] == minority_label)
majority_ds = ds.filter(lambda x: x["label"] == majority_label).shuffle(42)
return concatenate_datasets(
[minority_ds, majority_ds.select(range(minority_count))]
).shuffle(42)
if __name__ == "__main__":
parser = ArgumentParser()
parser.add_argument("model", type=str)
parser.add_argument("dataset", type=str)
parser.add_argument("output_dir", type=Path)
parser.add_argument("--lora-rank", type=int, default=8)
parser.add_argument("--lora-modules", type=str, nargs="+")
parser.add_argument("--num-epochs", type=float, default=3.0)
parser.add_argument("--batch-size", type=int, default=8)
parser.add_argument("--accum-steps", type=int, default=4)
parser.add_argument(
"--hub-upload-id", type=str, help="Name for HF model hub upload"
)
parser.add_argument("--token", type=str, help="HF token for private models")
args = parser.parse_args()
tokenizer = AutoTokenizer.from_pretrained(args.model, token=args.token)
tokenizer.pad_token_id = tokenizer.eos_token_id
tokenizer.padding_side = "right"
model = AutoModelForCausalLM.from_pretrained(
args.model,
device_map={"": torch.cuda.current_device()},
token=args.token,
# we can use bf16 if we're using lora because the base weights don't get updated
torch_dtype=torch.bfloat16 if torch.cuda.is_bf16_supported() and args.lora_rank > 0 else torch.float32,
)
ds = assert_type(DatasetDict, load_dataset(args.dataset)).shuffle(42)
train = balance(assert_type(Dataset, ds["train"]))
val = balance(assert_type(Dataset, ds["validation"]))
model_short = args.model.split("/")[-1]
def truncate_to_first_choice_token(statement, choice):
# We want only the first token of choice--this is where loss is computed
# Unfortunately the choice has to be encoded in the context of the
# statement bc of inconsistent behavior of some tokenizers (Llama, Mistral)
# So we duplicate work here, but it's fast.
s_toks = tokenizer.encode(statement)
full_toks = tokenizer.encode(statement + choice)
return tokenizer.decode(full_toks[:len(s_toks) + 1])
def format_fn(x):
lst = [
truncate_to_first_choice_token(s, choices[y])
for s, choices, y in zip(x["statement"], x["choices"], x["label"])
]
return lst
dataset_last = args.dataset.split("/")[-1]
total_steps = int(len(train) * args.num_epochs / (args.batch_size * args.accum_steps))
trainer = SFTTrainer(
model=model,
args=TrainingArguments(
f"{args.output_dir}/{model_short}-{dataset_last}",
fp16=not torch.cuda.is_bf16_supported(),
gradient_accumulation_steps=args.accum_steps,
learning_rate=2e-5,
logging_steps=50,
num_train_epochs=args.num_epochs,
optim=("adamw_torch" if args.lora_rank > 0 else "adamw_bnb_8bit"),
adam_beta2=0.95,
per_device_train_batch_size=args.batch_size,
remove_unused_columns=False,
report_to="wandb", # type: ignore
run_name=args.hub_upload_id, # for wandb
eval_steps=100,
save_steps=100,
save_total_limit=2,
warmup_steps=int(total_steps * 0.15),
weight_decay=0.1,
hub_model_id=args.hub_upload_id,
hub_token=args.token,
push_to_hub=args.hub_upload_id is not None,
),
data_collator=LastTokenOnlyDataCollator(tokenizer, mlm=False),
formatting_func=format_fn,
peft_config=(
LoraConfig( # type: ignore
r=args.lora_rank, target_modules=args.lora_modules
)
if args.lora_rank > 0
else None
),
train_dataset=train,
eval_dataset=val,
tokenizer=tokenizer,
)
trainer.train()